Andrej Karpathy shares his experience coding extensively with Claude in recent weeks, discussing how LLMs (Large Language Models) have impacted development workflows. Starting around December 2025, LLM coding capability surged, bringing a 'phase shift' to software engineering. The paradigm is shifting from manual coding to LLM agents handling most coding work while humans review and refine. Karpathy notes this change 'expands' development productivity and makes programming more 'enjoyable,' while also raising concerns about 'code quality decline (slopacolypse)' and deep questions about the changing role of engineers.

1. The Rapid Transformation of LLM Coding Workflows

Karpathy reports that his coding workflow shifted dramatically from 80% manual+autocomplete and 20% LLM agents in November 2025 to 80% agent coding and 20% edits+touchups in December. He now does most programming "in English, somewhat sheepishly telling the LLM what to code... out loud." While admitting his ego takes a hit, the ability to perform "large chunks of code actions" on software is enormously useful — especially once you understand LLM limitations and capabilities.

"I'm now mostly programming in English, somewhat sheepishly telling the LLM what code to write... out loud. A bit of ego hit, but the ability to take large 'code actions' on software is extremely useful."

This is the biggest change in roughly 20 years of programming, and it happened in just a few weeks. Karpathy estimates this change is affecting engineers in the double digits, though the general public is still largely unaware.

2. LLM Agent Capabilities and Limitations: The Need for IDEs

Claims that "IDEs are no longer needed" or "the agent swarm era has arrived" are still overblown, Karpathy notes. Models still make mistakes, and when working on important code, you must "watch them like a hawk." Errors are no longer simple syntax mistakes but "subtle conceptual errors like a somewhat clumsy, hurried junior developer would make."

The most common error type is the model making incorrect assumptions on your behalf and proceeding. LLMs also fail to manage their own confusion, don't ask for clarification, don't surface inconsistencies, don't present tradeoffs, and don't push back when they should. They also tend to over-complicate code, inflate abstractions, and fail to clean up unused code.

"Models still make mistakes. And if you have code you really care about, you need to watch them like a hawk from a nice big IDE next to it."

Despite these issues, LLM coding has brought enormous improvement overall, and going back to manual coding is unthinkable. Karpathy's current workflow: several Claude coding sessions in ghostty windows/tabs on the left screen, with an IDE on the right for viewing and manually editing code.

3. LLM Agents' Remarkable Abilities: Tenacity, Speed, and Leverage

One of the most impressive LLM agent abilities is 'tenacity.' Watching them persist without tiring or getting discouraged on tasks where a human would have given up long ago — that's when Karpathy experienced an AGI moment. This reveals that the core bottleneck of work was 'stamina,' and thanks to LLMs, that stamina has dramatically increased.

While it's hard to measure exactly how much faster coding has become, Karpathy feels he's doing far more than he intended:

1. He can now code all sorts of things he previously wouldn't have bothered with, and
2. He can tackle code that was previously inaccessible due to knowledge or skill gaps.

The effect is more 'expansion' than simple speedup.

LLMs excel at looping until a goal is achieved, and that's where most of the "AGI magic" manifests.

"Don't tell it what to do — give it success criteria and watch the model go. Have it write tests first, then make the tests pass. Put it in a loop with browser MCP. Have it write a naive algorithm that's very likely correct first, then ask it to optimize while maintaining correctness. Shift from imperative to declarative approaches so agents iterate longer and gain more leverage."

4. The Joy of Coding and the Changing Engineer Role

One unexpected change: coding has become "more fun" thanks to agents. The tedious busywork of filling in blanks has disappeared, leaving only the creative parts. Less feeling stuck or trapped, and more 'courage' knowing there's almost always a way to make progress with LLM collaboration.

However, others hold opposing views. LLM coding may divide engineers: those who loved coding itself versus those who loved building things.

5. Concerns and Questions About the Future: Skill Atrophy and 'Slopacolypse'

Karpathy confesses he already feels his ability to manually write code gradually atrophying. The ability to 'generate' code and to 'discriminate' code use different brain regions — you can still review well even while struggling to write, thanks to programming's many trivial syntactic details.

He expects 2026 to be the year of 'slopacolypse' across all digital media — GitHub, Substack, ArXiv, X/Instagram — and is bracing accordingly. He also warns we'll see far more "AI hype productivity theater."

Questions on his mind:

• What happens to the "10x engineer" productivity ratio? It'll likely grow much larger.
• With LLMs as weapons, will generalists outperform specialists? LLMs are far better at filling in blanks (the micro) than grand strategy.
• What will future LLM coding feel like? Like playing StarCraft? Factorio? Making music?
• How much of society is bottlenecked by digital knowledge work?

Conclusion

Andrej Karpathy emphasizes that LLM agents like Claude and Codex crossed a specific 'threshold of coherence' around December 2025, bringing a 'phase shift' to software engineering and adjacent fields. The intelligent part now feels far ahead of everything else — tool integration, new organizational workflows, processes, and overall diffusion. 2026 will be a very active year as the industry digests these new capabilities.